SkillRetBench

Dataset summary

SkillRetBench is a benchmark for query-to-skill retrieval: given a natural-language user request, systems must rank or select the correct procedural skill document(s) from a fixed library. The corpus is derived from a real-world production-style agent skill library (501 skills). The benchmark includes 1,250 queries with gold skill identifiers and optional distractors, spanning five evaluation settings (single skill, multi-skill composition, distractors, outdated/redundant, budget-constrained). Packaged baseline experiment results (BM25, dense, hybrid, LLM-style, and SADO-style runs) report standard IR metrics (Recall@k, nDCG@k, MRR, MAP) per setting.

SkillRetBench accompanies the SEFO framework (Self-Evolving Federated Orchestration) and targets research on skill-aware orchestration, large skill libraries, and retrieval under composition and noise.

Dataset structure

The release consists of three JSON artifacts (also listed in croissant.json):

File	Description
skill_corpus.json	Full skill index: metadata, corpus statistics, and the skills array (one object per skill).
skillretbench_queries.json	Benchmark definition: meta (counts, seed, generator) and queries array.
baseline_results.json	Aggregated baseline metrics: per-method, per-setting IR scores and a markdown summary table.

skill_corpus.json

• meta: generated_at (ISO 8601), skills_root, output_path, skill_file_count.
• statistics: total_skills, skills_per_category, description length stats, token_count_distribution, composition graph stats (skills_with_composition_references, total_composition_edges).
• skills (array, length 501): each element includes:
  • skill_id (string, canonical id)
  • skill_name (string)
  • description (string, short routing description)
  • trigger_phrases (string array)
  • anti_triggers (string array)
  • korean_triggers (string array; may be empty)
  • category (string)
  • full_text (string, full SKILL.md-derived body)
  • token_count (integer)
  • composable_skills (string array; referenced skill ids)
  • parse_warnings (string array)

skillretbench_queries.json

• meta: benchmark, generator, corpus_meta, total_queries (1250), counts_by_setting, rng_seed (42).
• queries (array, length 1250): each element includes:
  • query_id (string, e.g. SS-0001)
  • setting (enum-like string): single_skill | multi_skill_composition | distractor | outdated_redundant | budget_constrained
  • query (string; English and/or Korean natural language)
  • gold_skills (string array; expected skill ids)
  • distractor_skills (string array)
  • budget_tokens (number or null)
  • difficulty (string, e.g. easy, medium)
  • source (string, e.g. trigger_paraphrase)

Query counts by setting

Setting	Count
single_skill	400
multi_skill_composition	200
distractor	300
outdated_redundant	150
budget_constrained	200

baseline_results.json

• timestamp, meta: corpus_skills, queries_total, queries_by_setting, dense_backend (e.g. jaccard_fallback).
• baselines: nested object — top-level keys BM25, Dense, Hybrid, NaiveLLM, SADO. Under each method, keys match query setting names; each leaf holds metrics: recall@1, recall@3, recall@5, recall@10, ndcg@1, ndcg@3, ndcg@5, ndcg@10, mrr, map (floats).
• summary_table: pre-rendered markdown table (macro-style summary over settings).

Note: As documented in the artifact, NaiveLLM and SADO entries may be simulated (no live API) for reproducible baselines; see project script run_baselines.py.

Supported tasks

• skill-retrieval — Rank or classify skills given a query; evaluate with gold gold_skills and IR metrics.
• information-retrieval — Same retrieval formulation; comparable to passage/tool retrieval with long procedural documents.

Languages

• English — Majority of skill text and many queries.
• Korean — Present in korean_triggers, many query strings, and mixed Korean/English skill descriptions.

Dataset creation

Sources

• Skills were built from Cursor agent SKILL.md files under a single repository’s .cursor/skills tree (501 files), parsed into structured records plus full markdown bodies.
• Queries were programmatically generated by app.sefo.benchmark.query_generator with fixed rng_seed: 42 for reproducibility.

Methodology (high level)

1. Corpus build — Parse frontmatter and body; extract triggers, categories, composition edges, and token counts.
2. Query generation — Instantiate templates and paraphrases aligned to skills and settings (single vs. multi-skill, distractors, outdated/redundant, token budget hints).
3. Baselines — Run packaged retrievers/rankers (lexical, dense with configured fallback, hybrid, simulated LLM/SADO-style) and aggregate metrics per setting.

Considerations

Biases

• Corpus domain — Skills reflect one organization’s automation stack (dev tooling, PM, trading, cloud, etc.); frequency and vocabulary are not uniformly representative of all industries or locales.
• Authoring style — Descriptions follow a consistent “when to use / do not use” pattern; models may exploit format bias rather than deep semantics.
• Query generation — Synthetic and template-driven queries may over- or under-represent realistic user phrasing compared to production logs.

Limitations

• Scale — 501 skills is medium-scale vs. ecosystems with 10⁵+ tools; retrieval hardness may not transfer linearly.
• Gold labels — Defined by benchmark construction rules; edge cases and valid alternative skill sets may exist.
• Static snapshot — Skill text and ids are frozen for the benchmark revision tied to meta.generated_at in the JSON files.
• Baseline simulation — Some methods are noted as simulated in-code; compare only under the same harness and seeds.

License

This dataset is released under the Apache License 2.0. See Apache 2.0.

Citation

If you use SkillRetBench or this corpus, please cite:

@misc{han2026sefo,
  title        = {{SEFO}: Self-Evolving Federated Orchestration with Trusted Skill Governance and Skill Retrieval Benchmark for Recursive Agentic Systems},
  author       = {Han, Hyojung and {ThakiCloud AI Research}},
  year         = {2026},
  howpublished = {\url{https://github.com/hyojunguy/ai-model-event-stock-analytics}},
  note         = {Includes SkillRetBench: 501-skill corpus and 1,250-query retrieval benchmark}
}

Adjust howpublished / venue fields when a formal publication DOI is available.

Dataset repository layout (recommended)

After upload to the Hugging Face Hub, a typical layout is:

skill_corpus.json
skillretbench_queries.json
baseline_results.json
croissant.json
README.md

Additional documentation

• Croissant (ML Commons) metadata: croissant.json in this folder.
• Implementation and regeneration: backend/app/sefo/benchmark/ in the source repository.